A novel One-vs-Next approach for multiclass classification
Visham Hurbungs, Vandana Bassoo, Tulsi Pawan Fowdur · 2024
In machine learning, multiclass classification is the process of classifying data that contains three or more classes. One-vs-Rest and One-vs-One are two popular binary transformation approaches that reduce the multiclass problem into several binary problems. One-vs-Rest is usually faster and less complex than One-vs-One and is therefore the most commonly used strategy. The main contribution of this paper is the formulation of a novel One-vs-Next method for multiclass classification. The proposed approach consists in fitting one classifier per next class pair. This technique reduces the number of binary combinations and is expected to be faster since it requires to fit less number of classifiers compared to the One-vs-Rest approach. As a proof of concept, One-vs-Next was evaluated with the improved Scalable, Efficient, and Fast classifieR (iSEFR) designed for embedded machine learning on the Edge. In addition to its computational efficiency, experiments demonstrate that the proposed approach combined with the iSEFR decision threshold gives an average increase of ~20% in accuracy for a multiclass dataset with equal number of class instances.